SJC Land

Two-phase neighborhood evaluation

Assign every residential parcel to a neighborhood

Given that neighborhood structure

• Develop trending factors for each neighborhood with more than 3 or 5 sales

• Separate trending factors are estimated for land and improvement

Apply the trending factors and evaluate:

• The resulting ratio and COD

• The percentage of properties trended

• Repeat for each of the five years 2013-2017

Repeat for each classification scheme

• Current neighborhoods, townships, school districts, census tracts, etc.

• Further division based on clustering by acreage, year built, sq. ft., # of rooms

Describe tradeoffs between the different classifications tested.

2017 Sales

Initial Model

For an initial pass, we construct a model in which valid sales prices were predicted by land and improvement values. The values taken from this baseline model were then used to calculate updated assessed values.

With a very basic model, we were able to achieve reasonable values on both COD and Ratio.

Existing Neighborhood Results

As a next step, we computed individual regressions for each pre-existing neighborhood with at least 5 sales.

Min = 5

Min = 3

School Districts

Min = 5

Min = 3

Tax Districts

Min = 5

Min = 3

Township

Min = 5

Min = 3

Tract

In an attempt to find an additional way to create neighborhoods, we also used Census Tracts.

Min = 5

Min = 3

Combined

After examining individual geographical areas, we made an effort to nest geographical features into two levels.

School District & Township

Township & Tax District

Tax District & Tract

Clustering

Given the relative performance of tract, an effort was made to break the homes within tract into discreet clusters. The features selected clustering were acreage, year built, finished dwelling area and finished rooms. For each individual school district or tract, 2 distinct clusters were created.

School District Clusters

Tract Clusters

Mixed Models

Instead of fitting models to individual groupings, we also tried a mixed model in which every tract X cluster combination has its own slope within the model.

Previous Years

Given the reasonable results of the tract X cluster combination, we can see how it performs over the previous years.

2016

2015

2014

2013